Figure 5

Effect of positive contamination on classifier performance. Positive contamination, i.e. the fraction of positives in the unlabeled training set, affects the performance of a classifier. The figure shows two extreme conditions: a classifier trained with unlabeled data totally contaminated with positive examples (100%), and a classifier trained without positive contamination (0%). On the left the performance is shown in terms of AUROC (area under the roc curve), while on the right it is shown in terms of AUPR (area under the precision/recall curve).